Deep-learning versus greyscale segmentation of voids in X-ray computed tomography images of filament-wound composites
Publikation: Bidrag til tidsskrift › Tidsskriftartikel › Forskning › fagfællebedømt
Filament-wound composites (FWC) are prone to high void contents, with large and complex-shape voids. It is critical to characterise these voids accurately to understand their effect on part strength. The characterization depends on the accuracy of the analysis technique, for example X-ray computed tomography and the subsequent void segmentation. This paper compares conventional greyscale thresholding to deep-learning (DL) based segmentation. The processing steps for both techniques are discussed. The greyscale thresholding contains segmentation errors due to the simple one-parameter algorithm and the pre-processing operations required for segmentation. This reduces the accuracy of void characterisation. The DL-based segmentation is found to be more accurate for characterisation of void size, shape, and location. The processing-time and system requirements are discussed, helping to determine the suitable segmentation technique based on desired results.
| Originalsprog | Engelsk |
|---|---|
| Artikelnummer | 107937 |
| Tidsskrift | Composites Part A: Applied Science and Manufacturing |
| Vol/bind | 177 |
| ISSN | 1359-835X |
| DOI | |
| Status | Udgivet - 2024 |
Bibliografisk note
Funding Information:
The authors gratefully acknowledge SIM (Strategic Initiative Materials in Flanders ) and VLAIO ( Flanders Agency for Innovation & Entrepreneurship) for their support of the ICON project OptiVaS, running in the Nanoforce Program. M. Mehdikhani would like to acknowledge his FWO Postdoc Fellowship, project ToughImage ( 1263421N ). Abraham George Smith would like to acknowledge Novo Nordisk Foundation grant ( NNF22OC0080177 ) for their support and funding. The FWO large infrastructure I013518N project is acknowledged for its financial support of the X-ray infrastructure and the KU Leuven XCT Core Facility is acknowledged for the 3D image acquisition and post-processing tools ( https://xct.kuleuven.be/ ).
Funding Information:
The authors gratefully acknowledge SIM (Strategic Initiative Materials in Flanders) and VLAIO (Flanders Agency for Innovation & Entrepreneurship) for their support of the ICON project OptiVaS, running in the Nanoforce Program. M. Mehdikhani would like to acknowledge his FWO Postdoc Fellowship, project ToughImage (1263421N). Abraham George Smith would like to acknowledge Novo Nordisk Foundation grant (NNF22OC0080177) for their support and funding. The FWO large infrastructure I013518N project is acknowledged for its financial support of the X-ray infrastructure and the KU Leuven XCT Core Facility is acknowledged for the 3D image acquisition and post-processing tools (https://xct.kuleuven.be/).
Publisher Copyright:
© 2023 Elsevier Ltd
ID: 378176718